An ArcWelder for HP-GL: detects runs of polyline chords that approximate
circular arcs in plotter output and refits them into native AA (arc
absolute) commands. ~185 lines, stdlib-only Python.
Provenance: extracted in August 2026 from the working script that drives my Roland DXY-1150 pipeline — the pre-extraction development happened outside version control, so history accrues from here.
ArcWelder does this for 3D-printer G-code; this does it for 1980s pen plotters.
Vector pipelines (vpype, Inkscape, etc.) flatten every curve into hundreds of tiny chords. A modern machine glides through them; the 8-bit motion planner in a plotter like the Roland DXY-1150 executes each chord as a separate move — accelerate, decelerate, dwell, parse the next command over a 9600-baud serial link. On arc-heavy drawings the pen visibly stutters, plots take forever, and the dwell marks show in the ink.
The plotter has native arc support the whole time. arcfit.py gives the
flattened curves back to it.
- Parse the HP-GL, reconstructing absolute pen-down polylines
(handles
PU/PD/PA/PR; vpype's PU-absolute + PR-relative style). - From each point, greedily grow a run: fit a circle through the run's
first/mid/last points and extend while every point stays within
TOL(10 plotter units = 0.25 mm) of the circle and the angles advance monotonically (no doubling back). - Runs of ≥ 5 points become one
AA x,y,sweep;command; everything else is re-emitted as compactPApolylines. Rejected as arcs and left as line segments: radii aboveMAX_R(effectively straight), radii below4·TOL(circles so tiny the tolerance band is the whole circle), and centers beyond ±30000 (kept safely inside the plotter's ±32767 parameter range).
Tuning lives in the named constants at the top of the file: TOL,
MIN_ARC_PTS, MAX_R, MAX_CENTER, MAX_SEG_RUN.
python3 arcfit.py in.hpgl out.hpgl # in-place OK: same path twiceor as a module:
import arcfit
polys = arcfit.parse_polylines(hpgl_text)
hpgl_out, n_arcs, n_segments = arcfit.emit(polys)examples/demo.py flattens smooth curves into chords at 0.1 mm tolerance
(what a typical vpype pipeline emits) and runs them through arcfit:
| input | chords | output pen moves | reduction |
|---|---|---|---|
| 3 concentric circles | 149 | 3 arcs + 3 segs | 24.8x |
| 3:2 lissajous | 120 | 20 arcs + 18 segs | 3.2x |
| combined | 269 | 23 arcs + 21 segs | 6.1x |
A pen move is one PA coordinate pair or one AA arc — the unit the
planner dwells between. (Counted as commands on the wire it is tighter
still: the combined output is 30 commands — 29 drawing + 1 home move —
since PA moves batch into chunks.)
Bytes over the wire: 2776 → 855 (3.2x). The win scales with how circular
the content is — full circles collapse to a single AA, while curves with
continuously varying curvature (lissajous, spirals) weld into shorter arc
runs. Either way the plotter stops stuttering: one arc command replaces a
burst of micro-segment moves.
python3 examples/demo.py # regenerates examples/demo_{in,out}.hpgl
python3 test_arcfit.py # unit tests (stdlib unittest)This is the smoothing pass in a plot pipeline for a Roland DXY-1150 hanging off a Raspberry Pi:
SVG → vpype (layout, simplify, sort, HPGL) → arcfit.py → serial (9600 8N1, CTS)
examples/plot.sh is the real driver script (set PLOTTER_HOST, e.g.
pi@plotter.local); it copies arcfit.py to the Pi itself each run. The
~/plots/send_hpgl.py it calls is the author's own serial sender and is
not included here — any HP-GL sender with hardware flow control will do. examples/harmonograph.py generates a
damped-lissajous harmonograph SVG — a good arc-heavy torture test to feed
the pipeline.
- Only
PU/PD/PA/PRpen geometry is preserved. Everything else — includingSPpen-selects andCIcircles — is dropped from the output, which contains onlyIN;DF;plusPU/PD/PA/AA. Multi-pen plots must handle pen changes outside this tool (which is whyplot.sh --spskips the arcfit pass). - Coordinates are rounded to integer plotter units (0.025 mm) on parse.
LBlabel text is not skipped by the parser, so labels can yield phantom geometry (irrelevant for vpype output, which emits none).- Consecutive arcs chain without re-anchoring: each
AAcontinues from the previous arc's endpoint, so fit error compounds along a chain. Verified by simulating the emitted arcs back to points against the input: the worst radial residual against the fitted circle on the demo paths is 9.83 plotter units (0.246 mm) — insideTOL= 10, but near it. Reproduce withpython3 examples/verify_fidelity.py(also run in CI). TightenTOLand the chains tighten with it; an explicitPAre-anchor per arc would cost one extra move per chained arc, eroding the reduction. - Future: redundant
PU;PUpairs between polylines are a free ~3-command win, left unoptimized to keep the committed evidence files stable.
Physical before/after plot: coming after the next plotting session.
MIT